Loading summary
Transcript297 lines
- [00:00]
Jasmine
Oh my God. Someone should teach Trump how to vibe code. I think that would be so funny. Like I really want to know what you think of.
- [00:08]
Nathan
This is how.
- [00:09]
Jordan
This is how him and Dario are going to reconcile.
- [00:11]
Athena
Yes.
- [00:12]
Jasmine
Oh my gosh. Trump vibe code.
- [00:15]
Jordan
What app would Trump vibe code?
- [00:17]
Kimmy
ChatGPT and Claude's answers to what Trump would like to vibe code.
- [00:23]
Jasmine
That's a good etheld.
- [00:25]
Kimmy
Oh shit.
- [00:26]
Jordan
Ok, Claude. First one rally size AI uses satellite imagery in adv AI to generate crowd estimates that are always three to four times. Marshall says automatically generates comparison image showing his crowds versus Obama's. Oh my God, Jordan.
- [00:45]
Jasmine
I think it's funnier than us.
- [00:46]
Jordan
A CRM that tracks everyone he's ever interacted with and assigns them a loyalty rating based on public statements.
- [00:54]
Jasmine
Claude is funnier than we are.
- [00:58]
Nathan
ChatGPT are going to be bad. Did you ask ChatGPT?
- [01:01]
Jordan
ChatGPT still thinking. Oh my God. Tariff Master 3000. A calculator where you input a country and it outputs a tariff percent. The math is just math dot random times 100 plus 50. It has a reciprocal button that doesn't actually do math, but just adds 25%.
- [01:22]
Jasmine
No.
- [01:23]
Jordan
Okay. ChatGPT is not funny. Kimmy's also not funny at all. Okay, Claude, you have the mandate. Congratulations, Dario, you did it. Hats off to you. Now for Putin, we've got to be rough. Long table augmented reality that automatically previews how small visiting leaders will look from him 40ft away. Historical Russia AI you draw any border on a map. 3000 word justification for why that territory is historically Russian.
- [02:01]
Nathan
Oh my God.
- [02:01]
Jordan
Oligarch Manager. A portfolio tracker, but for oligarchs track their yacht via satellite monitors for disloyalty signals. Has a one click widow alert feature.
- [02:12]
Jasmine
Jesus Christ.
- [02:13]
Jordan
Okay, Tucker Booker. An app for scheduling friendly interviews with contrarian western media figures.
- [02:22]
Jasmine
No.
- [02:26]
Kimmy
Tada. Overfit. I think Nathan wanted to kick us off.
- [02:31]
Nathan
People need to use this shit. The coding agents are remarkable. They're going to get a lot better. I think. I do everything with them. I'm like way more. I'm going back down the stack as a research manager into more like legit engineering tasks. I manage all my book edits. I'm just like a fanatic. Anything can be taken on with these. And I also think with the insider perspective that these tools were only proposed last about April by people not on the core modeling team. So as the people on the people, as the people who build the models get more into these, the models are going to take big steps at being able to use These tools. I still think we're very early on this cycle kind of in combination with the fact that the models actually are getting better at a very steady pace this year. I think this is downstream of the whole inference, time scaling, tool use, thinking model stuff from last year. Like the organizations are in a spot where they can hill climb on all of these methods from pre training to post training to tool use so well that it's going to get so much better. And the value unlock is hard to measure. People say like how many hours a week of work are you getting out of cloud code? And it's just like how do I put a number on stuff that I would just never do because I wouldn't fit the time into my day. So I really think that I'm changing how I work and I want to take time to understand what's the right way to decide what to work on now that I have way more ability to be dynamic in what I work on and have way more output. But I just think that it's going to be a very weird year as people kind of get through this.
- [04:16]
Jasmine
I was talking to a friend this morning who is a startup founder and was doing some engineering interviews and he said that one of his engineers this morning used the phrase Claude code psychosis. And then he was like, did he get it from you? And I was like, could be transistor radio, we don't know.
- [04:33]
Athena
I'm most curious about like what's the maximum number of agents have you used at the same time.
- [04:40]
Nathan
For real work too? It's like I can't multitask supervising them in more than two things. Send off a Claude on YOLO mode to see what it does. But most of the things that I'm doing with I actually care, like editing my book, like I have to provide feedback on literally every single sentence or like research code. It's like it's not gonna like I could send it off to go build a website for me, but I'm not doing that type of work. Like this is just like turning through my to do list. So maybe I should get better. Maybe it's too low. I need to like take time at the beginning of the day to chart out all my clods.
- [05:20]
Athena
I feel like the ability to mobilize theoretically an infinite amount of sub agents to work on things for you is like a giant enabler of add on steroid. Like I've now tried to have six sub agents to work on a website for me and it's like by the time I'm telling Agent six, what they're doing. I've like, gone to the habit of not worrying about what I've already told subagents 1 to 5 to do, but it's still. It just like, makes my brain feel so good that after one prompt, like, everything from my brain's emptied out and I'm just like, waiting there for the fastest of the six sub agents to get back to me.
- [06:05]
Jasmine
Wait, Athena, can you explain, like, what are subagents and what do they do? I. I think I use subagents once on accident, but like, I have not quite figured out what they are yet.
- [06:17]
Athena
So I had to look this up on claude. Not Claude code for CLAUDE to explain to me how subagents on CLAUDE code do their work. But essentially, if you give Claude code a prompt and it's a relatively complicated ask, you can hit Control B for it to work in the background. You could also specifically say in your prompt, hey, I want to spin up some parallel workflows, and I'm going to have agent one work on ABC and Agent two work on cde. Where I wanted to nudge this conversation to is like, agents and human developers don't behave the same. Before a human had the ability to spin up like six up clones of my stream of consciousness, I couldn't think in a way that is like, in parallel. I would only think in like a. A chain of thoughts type of way. But now with the ability to. To just mobilize an army of sub agents, like, I am changing my beh to accommodate how I'm enabled on these codes and platforms. And that is really cool and I think will have lasting impact not just on how technology products evolve, but also how we, as the human behind it or the labor force, interact with the technology. So I'm gonna go have fun.
- [07:35]
Kimmy
That was such a good take.
- [07:38]
Athena
Just cut that out.
- [07:39]
Kimmy
I'm not cutting it out. We're gonna leave it in. That was a great take. All right, have fun.
- [07:43]
Nathan
Applaudees.
- [07:44]
Kimmy
See you later.
- [07:45]
Nathan
The best part is, like, it is a great take and it has like a little bit of echo in it, but you just gotta leave it in.
- [07:50]
Jordan
I think.
- [07:51]
Kimmy
I think Athena's point is fascinating because there have been sort of technological transformations that have fundamentally changed the way you think when you work, right? Like, going from being able to. From having to handwrite to being able to type three times as fast, that, like, allows your mind to work in certain ways. Being able to Google anything on command, like, allows your mind to work in certain ways. But this new thing of like, Just need to like come up with one sentence to tell a sub agent to do a piece of this project. Well, maybe the question to to Nathan is like, is this going to be something that sticks around or are the models just going to be so good that this is just like a, an interesting like interstitial moment?
- [08:36]
Nathan
We could think about some of the failures that we're seeing and they're definitely not perfect at staying on task even yet, like even now. So the things that at least in the ML coding I'm doing, I see them do stuff and I'm like how did you even get that idea from a fairly simple prompt? Like I give them a fairly simple prompt and because I think the task is simple and then they just start like modifying the wrong script and I'm like how do you even get that idea? Some of those things I think will be tampered down. But the crucial thing is that we're going to start using these on more complex systems where the specification really matters a lot. So therefore most of the work is going to be done in maintaining the specification. So I've been talking about how I set my CLAUDE up on a DGX Spark, which is essentially like a small Blackwell GPU that you have on your desk that's set up to be like a, essentially like a Mac Mini but Blackwell gpu. And I have different levels of CLAUDE directories where I have the actual code base which is on GitHub underneath a higher level CLAUDE which has access to my setup docs on how you use this computer doc where I have CLAUDE dump every single command it runs to launch an experiment. A high level Claude MD with how I am approaching using this computer and I, I, those things are not going to go away. It could be that you really don't have to iterate them on them as much. That's the kind of thing is like how fast did they become in making progress? It's, I don't think the one sentence prompt is actually going to be that lasting. It's just so nice when you like are just trying to fire something quickly off. I just think it's going to get complicated. But that doesn't mean the point about running six agents in parallel and thinking about how you work is different. I think that the lazy thing that I do is I'm like claude, write me a prompt for GPT Pro explaining the situation and then GPT Pro writes a detailed report with links to open source repos and hundreds of words on the specific issue and it's like I'm Doing that because I could write this prompt, but it would take me like an hour of hard work and I'd rather just wait for GPT to do this. And I could link these, I could wire them. I just don't want to have. I have a ChatGPT subscription. I don't want to spend on the API. But I set up the equivalent of Gemini, which is like I'm using Claude to generate technical diagrams for a book and I'm just like, it's not that good and I don't want to give it feedback all the time. So I have a Gemini feedback skill. So I'm just like, go talk to Gemini. And it's pretty good. It just reduces the number of cycles I need. The high level thing is back to this. There's two points of will we actually need to document. I think we will if you look at GPT Pro. And some of these models actually do a lot better when you give them a really clear specification. So I do think people will start doing that more as they start. I call it as how do you stack bricks rather than just making toy apps for making productivity apps or whatever. Athena asked the same question that I'm wondering is like, how do I properly allocate my time as they get more autonomous? They're not super autonomous in many ways now. So we have a bit of leeway to think about it, but it's surely going to come where we don't have to provide that much feedback and it just iterates. I don't know. I've rambled a long time.
- [12:11]
Kimmy
We should take a detour to our fun toys we've all built. Jasmine, you want. Why don't you start?
- [12:18]
Jasmine
My high level. My high level thought on what Athena said and what Nathan said is that I think that it is true that Claude code and most of AI we have had so far just turns everyone into a pm. I used to be a pm and so having subagents is like having six engineers. So it was very. It's very funny because it scratches the exact same itch as when I had a team of like 12 engineers and then I would literally on slack fire off 12 different commands and then like they'd go off and do their thing. And then the other, I'd be like, slack Notification engineer wants to know what to do about the sedge case. Slack Notification engineer says it's done, can you please test this? And it feels exactly the same to me, except they're all, I guess, the same engineer. So actually I was like, oh, this is Just like it is a very sort of add way of thinking and it's very distracting. But it was very reminiscent of my PM days in a way that I kind of missed and. Yeah, and just like at a, you know, at a company, if you're at a small startup subsec was like on the smaller end, then it's okay to do a lot of small tweaks, one line. But as soon as you want to, you know, build a more complex feature, you do. You actually have to go write a spec and work with engineering manager and plan it out more concretely. But yeah, I think all of this is sort of coherent with my view that Claude Code just means everybody's a PM now in terms of toys. Okay, My most recent toy that I'm excited about is my imessage wrapped. I styled it and added a bunch of stuff. Maybe I can send some screenshots later so that you guys can put them up. But basically I started doing this because I was thinking about how, you know, like engineers and programmers and researchers have all these like software shaped problems and cloud code can go like, explore their code base and like build things and it has all this context and it's so useful and like my life is in mostly Google Docs and like text messages or like it's more relational and like thinky work. And I was trying to think, you know, like I could say, oh, AI will never be able to do the relational parts of my job or the PM job. But like, what if it could? What if I also gave it like the right training data, the right context to explore. So what's the equivalent of a code base for my relationships? I was like, okay, maybe it's iMessage. And so I basically wanted to see, could Claude Code 1 make an iMessage rep? Because I thought it would be funny to see just how my relationships have changed over the past eight years, nine years, whatever. But two, could it actually then also understand what people's relationships to each other were and do things like I did? In my head it was dinner party evals. So I made three groups of people from my text who it thought would get along or be like, these are the people you are more emotionally sharing with. This person you seem really guarded around and it would like give reasons why. And so basically I was trying to understand, does Claude have emotional and relational intelligence? Because if you imagine how does it keep climbing the ladder of abstraction in an organization and move up from say a low level software engineer to a low level PM to like a mid level pm you start to need to learn, like, these skills and like, you slack data and, like, whatever. And so anyway, I, like, learned a bunch of things. Like, I said AI 43 times in 2020, and then, like, or 43 times in 2018, and then like 79 times in 2023 and like 700 times this year or something like that. So I have all of these funny trends or like, this person use really good grammar with and this person you're apologizing to all the time and lots of stuff like that. So I've been having a lot of fun.
- [15:33]
Kimmy
The trend analysis reminds me of, like, people charting she speeches over time, being.
- [15:38]
Jordan
Like, oh, new forces of production had a rough year. Like, oh, we didn't mention corruption. Is that. Is that bullish or bearish?
- [15:47]
Jasmine
Yeah. Oh, my God. The most embarrassing one was it tracked. It was like, I can see the crypto bubble in your texts because in 2021, I said the word crypto like, 400 times. And then it went to, like, two the next year.
- [16:01]
Nathan
I feel like Claude Code likes a snarky statement like this. It's like, my stuff. It loves to be like, oh, fascinating result. It's like, definitely trying to, like, sycophant you into thinking that you, like, did something good with it.
- [16:14]
Jasmine
I felt called out. I don't know.
- [16:18]
Kimmy
All right, should I. Should I do.
- [16:19]
Jasmine
Go for it?
- [16:20]
Kimmy
Should I do mine?
- [16:20]
Nathan
I had a few.
- [16:21]
Kimmy
First, we made the Jewish themed mahjong trainer for my mother. We bought a mahjong set. We gotta get her up to speed. There are a lot of hands to learn, a lot of tactics, and it has little Yiddish jokes throughout. I think the best one I done were all these kind of, like, dialogue ones where you put it through. You know, I have like, an API to 11 labs, and then I have another API with Hume, and it kind of like, listens to your speech. Like, it voices you back and then analyzes it. I did an Acting Trainer 1 where you're, like, doing a table read or you're doing improv, or you're, you know, reading a Shakespeare speech. And you give this speech and like, if you're low energy or you don't hit the note you want, then it will, like, Hume will say, oh, this was like, his voice had a lot of anxiety. Like, maybe like the. Like, it's like as if a director is standing there being like, here have, you know, say this line with more confidence, which I found fascinating. I have the relationship trainer one where when someone is not supposed to, like, if someone is having a tough Time. And they want to just be told, like, I'm sorry you had a bad day, not do like the fucking boy thing and try to fix it. You get like negative points for trying to fix it. So.
- [17:49]
Jordan
Yeah, a lot of.
- [17:50]
Kimmy
Lot of fun things I've been up to. Oh, the best one, I think, is for our one and a half year old. I made like a keyboard where every letter you press is like. B is for like, a little thing pops up. B is for Ball, A is for Apple. And I set up a thing so it wasn't the robot saying it. Like, I recorded it. And then I had a Chinese version and then I had a weapon system version. No, but I did a one to her. I felt like that was a little much. W is for Warthog. We'll keep to the version of that.
- [18:29]
Jasmine
Before she gets to be a guest on Second Breakfast.
- [18:32]
Nathan
No. Yeah, it's like you're making a little bumper for sound effects for Second Breakfast where like, one of the other guests could press their keyboard and then it's gonna be like, D D W's for war. Like insert the Jordan sound.
- [18:45]
Jasmine
That's really good.
- [18:46]
Kimmy
I'm sorry, Nathan. Aside from like, I don't know ML shit you've had, what fun of you.
- [18:53]
Jasmine
He's so productive with his quad code.
- [18:56]
Nathan
I just have an insurmountable to do list, so I'm just like, trying to crank everything I have. Wedding planning, love. I've only done wedding planning in like. I didn't use Claude code. I used like GPT Pro to write a bunch of Excel formulas back in the day to figure out, like, who we classified our friends into. You gotta use, like, where. Where they're coming from if they're going to fly or drive. And, you know, everyone ends up with like a range invite list now. I guess I'm almost there. I have all the addresses in an Excel sheet. I'm unstoppable. Could send them stuff. I could ask Claude to buy the certain people gifts if I think they're falling out. You know, rate my friendships.
- [19:38]
Jasmine
That's amazing. That's really good. I am actually. Actually, yeah. Next time I go to buy someone a birthday gift, I'm going to have Claude code recommend me some birthday gifts to buy based on all the texts. Personal.
- [19:50]
Jordan
Yeah, it's like thoughtful on demand.
- [19:51]
Kimmy
That's great.
- [19:52]
Jasmine
Personal CRMs are out. Having cloud code do things for you with access to literally all of your data is.
- [19:57]
Nathan
It's like, oh, how did you remember that?
- [19:59]
Jasmine
That was so thoughtful. It's because you mentioned it five times. In February 2023 I downloaded then it sucked.
- [20:06]
Kimmy
I used it for like a month. But it was more a business CRM.
- [20:09]
Nathan
Than a personal one.
- [20:10]
Kimmy
I was like uncomfortable with the personal part.
- [20:14]
Nathan
Who of us 3get gives Claude her credit card first?
- [20:17]
Jasmine
Jordan.
- [20:18]
Nathan
Jordan.
- [20:21]
Jasmine
I think it's Jordan.
- [20:22]
Jordan
Oh, oh. The other one I set up was.
- [20:25]
Kimmy
I was trying to find the whale bets on polymarket. Initially it was called Polymarket Tracker and then I was like Claude come up with a funny name for it. Calls it Polysnitch.
- [20:36]
Jasmine
Oh, that's good. That is pretty good. It's really good. It's really good. Are you becoming any more productive, Jordan? Like have you found actual work use.
- [20:44]
Nathan
Cases does sort of work here. Just reads and yaps.
- [20:48]
Kimmy
Well actually Irene, Irene, Irene has a whole article on this of like Claude using Claude Claude for China analysis. It can, it can pull from RedNote and WhatsApp. Excuse me. It can pull from RedNote and WeChat.
- [21:02]
Jasmine
Oh, that's pretty cool.
- [21:03]
Kimmy
They have APIs and like you know, doing text analysis on. On that sort of thing is going to save folks a lot of time and be really cool.
- [21:11]
Jasmine
Wait, is this on this coming out soon.
- [21:14]
Kimmy
Coming out soon.
- [21:14]
Jasmine
Okay, okay, okay. Yeah. Because I am illiterate and so this will be great for the illiterate China watching community.
- [21:25]
Nathan
The other interesting thing is like how do you think more people that aren't dealing with computing heavy jobs get onboarded? I think for me it's like being around AI and CS people. It's so obvious. It's like I write my book in overleaf, which is secretly a git repo. So Claude can go through and find literally all my editor comments and be like, in chapter 14, your editor said this. What do you think? It's like I just literally can do my entire book editing through Claude and that's because I'm used to these environments and it's like, it's like I almost think the book editing world should get there, but I don't expect them to for at least a decade. But there's a lot in between the middle of that of just like, I don't know if maybe the multimodal stuff will make it so people who have like can always make the exact way to use a computer that they want or something. But just getting a better sense of what other like knowledge work is I think would be pretty useful to contextualize what's going to happen this year.
- [22:25]
Kimmy
Is this time for the overfit mother segment.
- [22:29]
Jasmine
Oh yeah.
- [22:31]
Nathan
Hi mom. I know she doesn't listen. My mom asked if I could say hi mom, the next time I go on a big podcast like Lex.
- [22:37]
Jordan
Sorry, I'm not sure this counts.
- [22:40]
Nathan
I don't know.
- [22:41]
Jasmine
China Talks, a big podcast. The number one China podcast.
- [22:44]
Kimmy
Hey Jasmine, how's. How's your mom been? Cloud coding?
- [22:49]
Jasmine
My mom has been cloud coding. She made two things that I know of. One is she, I think tracks her workouts in like an Excel spreadsheet normally. Like she says what she does what days, like what kind of training, whatever. And she has moved that tracker to like a web app that she did with cloud code. And then the other thing is my mom does a lot of like ebay reselling of like buying and reselling like old clothes or old things. So she made like a revenue cash flow type tracker for like all of the items, all of the revenue. It does like she has to like export her data from ebay because I think there's no API for it or whatever. But she has that now and now she can sort of like visualize all of those things in her projects. I think it's really cool because my mom was a software engineer back in the day. She studied computer science in like 1990 and then she like left the workforce probably around 2000 or so basically to take care of me and my sister. And so she felt, feels like not that familiar with modern coding and modern programming, but she really likes software and she wishes that she did more of it. And so I think Claude code has been a very fun way, I hope fun for her to get back into building software again in a way that is less intimidating than it previously was to try to onboard to all of these new languages and frameworks that she just missed by a couple decades. And so our family group chat is me and my mom sending Claude code projects to each other. But we haven't gotten my dad and my sister on board yet.
- [24:18]
Nathan
I haven't even gotten my brother and he works in tech. I text him like every couple weeks. I'm like, bro.
- [24:26]
Jordan
Yeah, it's interesting.
- [24:27]
Kimmy
I wonder what percentage of this audience has listened yet or has. Has. Has done it.
- [24:34]
Nathan
Guys, can we put a form in the show notes like, like, like just like one question. Really easy. Yes.
- [24:42]
Jasmine
I don't think so. This is one of my takes. It's not easy.
- [24:45]
Kimmy
I.
- [24:46]
Jasmine
Sorry. Well, no, it's not easy if you're really non technical. Like it's obviously easy if you're knowing.
- [24:51]
Kimmy
What terminal is that.
- [24:52]
Nathan
That is like if you know what a GitHub repo is no, but most people it's going to be easy. But yeah, that's what I'm saying. Like if people don't know what a GitHub repo is, I don't think it'll be easy.
- [25:01]
Jasmine
Yeah. So like this is actually one of the things that has been really bugging me about the cloud code discourse is like I think that like technical people are really non calibrated as to like what's easy and what's not easy in the same way that I'm sure like I'm uncalibrated about things in my domain. But like, like I had never touched my terminal before. Like I worked in software but like I don't have to touch my terminal when I work in software. I just like slack and engineer what to do. Um, and so, or like when I was working with my mom for example, she at first she tried to type into like normal Claude because like how do you even know that Claude code is not just Claude, right? Like she's like well I told it to code. Like that's what you think it means, you know, so there are things like that or like the difference between what is local versus what's in the cloud, right? So it's like okay, like now you have this app but it's on localhost, like how do you make that a real website? Like people don't know that. Or I was stuck on again like I never used my terminal before. I was like why can't I click around and why can't I backspace normal? Like there are all of these things that you have to onboard onto and like if you just stick at it for like an hour or two as a non technical person you will figure it out. But I think there's like all these little steps that are just like way harder than people think. Or like every time I ask for a permission it'll say a bunch of things that like you don't understand. Like if you're a non technical you don't understand what it's saying and like you might make the mistake of looking up what each of those are and then you get super overwhelmed. But actually the answer is to always approve. But like why would you do that?
- [26:22]
Nathan
You know they're going to make it an app because installing it is not easy. Because if it's an app they'll be a lot easier for weird clicking around. And also just like you know, the ChatGPT Pulse thing, they're probably going to make it an app with an onboarding flow that's like how intimidating is this? And then just kind of slowly work you down the stack because if you have somebody telling you how to do it, it's really not. The mechanics of it aren't hard, it's just not like people definitely be like oh my terminal's open. What the hell is this thing?
- [26:54]
Jasmine
Yeah, it looks hackery. It's scary.
- [26:57]
Kimmy
Well, so has gotten anyone tried Claude cowork yet? Because I guess that's a version of their answer.
- [27:02]
Nathan
I haven't. No, I just heard it was like a dumbed down version.
- [27:06]
Jasmine
That's what I heard.
- [27:06]
Nathan
But it's like. But the interface is probably better in that regard.
- [27:11]
Jasmine
Yeah, but one thing I was talking about with friends recently is like, you know, every once in all those polls come out and it's like what are people's like normal people, Americans level of familiarity with different AI companies and like OpenAI is like 80% and then Anthropic is always like 8% and it's the same amount as a made up AI company that they use as a control variable. So they'll make one up called like like you know, deep AI or something. And Anthropic always ties with the one that is like literally not a real AI company because people don't know what it is. But perplexity is actually really high. And so I was talking with some friends yesterday about why are normies so into perplexity when like nobody I know in SF who's like serious about this stuff would ever use perplexity. And like we don't know exactly but like they invested a ton in paid acquisition and advertising and creator partnerships, especially on universities. The interface looks like a Google search. So if you don't understand chatbots and like a chatbot is like scary and weird to you. The fact that Perplexity is like kind of a Google shaped wrapper actually just like does a lot psychologically. And so I do sort of think that like there's just like a. When normal people don't even know what normal Claude is like there is a very long way to go before they're comfortable with like Claude code or anything coding agent.
- [28:22]
Kimmy
Well it's interesting because we're supposed to.
- [28:23]
Nathan
Be, it's going to be tech centric on the cloud code thing. Yeah, I don't think cloud code adoption is going to feel that sea change end of the year like it's still going to be like I think Anthropic is happy to be in this more business selling because anyone that's like in tech, even if it's personal projects is kind of like selling to a business. Part business market market. So that, that I don't actually have a good way to predict what that will mean for it. But I would guess Anthropic's happier about it.
- [28:53]
Kimmy
Who's doing super bowl ads?
- [28:55]
Nathan
I asked my team if we could do an Atom Project super bowl ad and I got shut down. I don't have the money for that.
- [29:02]
Jasmine
Do you think that the super bowl is your primary audience demographic for the Atom Project?
- [29:06]
Nathan
It's a big audience. This is a. This is the type of idea that I have downstream of having worked with Claude too much. I'm just like, why can't I do psychosis?
- [29:17]
Jasmine
That's like, you're too high agency now, Nathan. Like, you can't do that. That is too much delusion.
- [29:24]
Kimmy
I think it seems like, I mean, they're starting to do mass marketing. I mean, IPO will be an interesting moment for like press for them. But yeah, it's like, look, if they're making all this money now, do they. Do they really need it? Like, it's, it's. I mean, we spent, we spent months talking. I mean, we spent years even being like, like everyone. I mean, the transistor radio energy around Claude, around Anthropic was like, oh, they're.
- [29:51]
Jordan
Feeling good, they're cooking up some stuff.
- [29:53]
Kimmy
They think they're hot shit. But when you were talking about revenue and user numbers, it really wasn't anything competitive. But I don't know, Nathan, what's the competitive response here from opening hour? Are they cooked?
- [30:05]
Nathan
Codex has a very different nature where it is far less fun, but it feels very capable. I think especially if you use the really high reasoning efforts and you've given a very, very specific specification with test cases. So there are people that are very, very team Codex and potentially in a way that's like both meaningful engineering and more asynchronous. Also, Sam Altman has had a classic Sam Altman tweet today where it was like, big updates to Codex coming. We're going to have to take cybersecurity really seriously. And I'm very torn where I'm like, I hate how he talks. And it's making something that should just be a disclosure kind of into a hype post. But probably directionally means that Codex is also getting better. But if you use both of them, the nature is very different. Claude feels like a video game and Codex feels like a kind of hard to use super power tool. So the knowing people that work in tech feeling like A video game is actually probably more important because people like to have fun more than they like to get shit done. So I'm like, very polished. I mean, that's why I'm like, I just like this. I'm like, yeah, gotta go check on my cloths, man. Yeah, they've been alone for too long. Like, what the heck? It's like. Like, that's a very real sentiment. It's like the sick. The types of things like circumcy and model behavior are very different in the agent interface. And it's interesting to me that anthropic has kind of hit it well again here where it's just like what the model actually outputs in terms of the tokens. The vast majority of it we never see. So the model could do such weird stuff. But, like, previously, if the model just outputs a bunch of crap, you're going to be like, I hate this model. So there's a big sea change in that that I'm personally very interested in. For researcher.
- [32:07]
Jasmine
Yeah, it's interesting because I feel like the. I'm like, extremely bullish on chat as an interface in general. Like, I think chat is the best interface because that's how people talk to each other. I'm bullish on voice too, but just like conversational, I think, is like always the correct thing. And it's interesting because with ChatGPT or whatever, chat is like, it is the product, it is the output. But then with quad code, chat is just like the interface. It's like there's like this scary machine of like crazy shit happening that I don't understand. And then there's like the chatbot and it. Like it's the smiley face on the shogoth, right? Like, it's like me talking to the chatbot and it sort of abstracts away the scary stuff in a way that I appreciate. One thing I'm curious about, Nathan, is like, so why is Claude code so good? Like, how much of it do you think is the U? Like the command line UI? Is it the model? And like Opus 4.5 being better.
- [32:55]
Nathan
Like, most of it is like robust tool use. So the model is very good at calling. It's like it very rarely messes up the shell commands that you have to run to move files around and run things there. Like, knows what to run. So you kind of abstract most things into some sort of code or computer Systems problem. When GPT, when Opus 4.5 was released, it had a pretty interesting profile and a model which was that Opus is historically more expensive, much more expensive. Than other models on the market where they, they did a price cut per tokens, which I think is down to them getting better infrastructure so it's easier for them to serve and the model matching the infrastructure better. But then also the model's response length was way less than Sonnet. So like with this reasoning, models use a lot of tokens. So Opus 4.5 was very efficient. And then the other thing is that these bigger models tend to be a bit more resilient and weird. Use cases where they don't break and output weird tokens, which I think makes it so that Opus 4.5 is like a lot more robust in getting the right completions. It uses fewer tokens to get there. So like, all of those things combined into an already good interface just made it much, much better. There's a lot of talk that, like, oh, the, the interface actually got a bunch better like in December, well after the release of the model, which could have been like, it's some weird prompting thing or some change that makes it a bit more robust. And that's the really interesting thing for me is like, how much of this is actually a harness that they did really well. But I think, like, Opus definitely suits this where it's like, it's just easy to mess stuff up. But these models, like, you see them mess something up, it's like they mess up a git command and they have to redo it. And that just adds a lot of latency. And I don't think people like to see, even though we're all vibe coding, we're throwing things out there. I think people kind of like to see the model making progress. So I have a lot of like, I'm not super sure on what's happening, but the actual Opus model release does have some hints that the fewer tokens is so nice. Models that take a long time to do something is actually like, really annoying and you have to work differently with it.
- [35:18]
Jasmine
Yeah, I feel like the GPT thinking models, I mean, they've gotten a little bit better with 5.2, but like, like they very frequently take forever to do, like a very simple thing and like the ratio of like, how long it takes to like, the quality of the output or like, you can just tell it's taking like the least efficient route possible to do something and then like, spitting out so many words that, like, I don't care about or, like writing unnecessary scripts and whatever. And yeah, I do feel that Claude code is. It feels even though it's annoying to like, wait for it. Like it Feels relatively efficient and I don't know what that means but like it feels that way relative to using a GPD thinking model right now.
- [35:52]
Nathan
Know even I get weird things where like codecs if I don't turn the thinking effort up. Like sometimes it just struggles with basic git operations which like Claude's been able to do for a while which is like check I I often will start where with an agent where I'm like check out a new check out main reset to or reset to the remote which is like reset to the the head of the tree which is the latest stuff make a new branch and it's like sometimes codex messes this up at least on normal settings but when I upgrade to like extra high it can do all these stuff with more reasoning effort. So I don't know if their model is just still a bit funky or something there where Claude's just designed focused on these easier things. But I would say that more of this is we're still very early on it. Rather than thinking of these as lasting patterns in the space. I think there's a lot of evolution that'll happen.
- [36:46]
Kimmy
Nathan, where does this leave all the Chinese models?
- [36:48]
Nathan
And they have their own coding plans. I haven't tried them. Like I'm a sucker for intelligence and performance and I'm willing to pay if something is better. But they are. It's an interesting niche where everybody's like you have to pay 20 or $200 a month for these top of the end models and they're like pay US$2 a month and we'll give you one of these things. Which all in is probably a much better position than they would have expected to be in to. There's like I think Z AI actually has meaningful adoption on theirs and I 60% confidence statement is something like they had to pause signups for new people or something. Something along that lines of like where they don't have the supply to serve all of their GLM coding plans or something like that. So like people are using these. I mean Minimax was subsidizing their model to be free in an open model focused CLI agent open code. And then that that trial like ran out And I saw people on Twitter, they're like damn, like this is no longer there. Like these companies are focusing on and getting a lot better at navigating the western media ecosystem and having influence for their AI products. But if you just think about it like the value of having a model and a harness that you can run on a really beefy MacBook to do a 10% of the performance of Claude code is very high. Like that stuff is maybe not end of year but in the next few years that is a real thing. You on an airplane will have a model. I will have a model on my laptop within a few years that is as good at these coding tasks as Opus 4.5 is now. Just by making a model that's really, really tuned to it. I really think the local models are going to make strides there won't do as hard of tasks but, but that's a really interesting market to have influence over.
- [38:42]
Jordan
Is it though Nathan?
- [38:43]
Kimmy
I mean how often are we on airplanes? Like I, I, I actually don't understand.
- [38:47]
Nathan
People don't want to send their model. People don't want to send their information to the cloud, they don't want to send their information to AI companies.
- [38:56]
Kimmy
Really? I don't know. But my, my, my base case on this is like the market share of Brave. Which is what?
- [39:03]
Jasmine
Well I guess it depends who people are, right? Like are we talking like normal consumers or are we talking businesses? Cause like regulation is like one of the big ones, right? Like if you work at a bank, you can't use ChatGPT, you can't use cloud code, you can't move your data anywhere just because regulation or like same thing with defense and govtech companies. And so part of the local stuff might just be like regulated spaces.
- [39:24]
Nathan
That's how we think about it. But I mean it's not going to be a majority of the pie. I just still think it's interesting in terms of like you effectively don't pay marginal cost, you fully control the technology. Like I just think it's very interesting in that perspective for education for example, like people will still use these like a university, it's going to be much happier to tell everybody to download something off the Internet for free than sign all their students up for A$20 a month coding plan. Like I just think that it's just like the history of people use open source software that's not as good for a variety of reasons. And there will be a Venn diagram that is overlapping but different for AI. And it's new that it's being dominated by Chinese companies. Like yes, this is not, I still don't think it's like the top end story of AI generally like the like open source AI influencing research and long term R&D across US and China is probably a bigger open model story than like local coding agents. But I still think that all of these battlegrounds are good for Learning about AI dynamics and will people actually use this and stuff? Yeah, but it's not.
- [40:45]
Kimmy
Most of them aren't majorities or one attic data point. Afra, friend of the pod, just had her wedding in China and was chatting with folks. Every single one of the developers, like, had found their VPN way to get at quad code because they can't be bothered with like the, you know, 30% less smart but subsidized mini Max substitute.
- [41:09]
Nathan
Isn't it's so funny because didn't Anthropic tout how they're like banned in China and they went through all this effort to ban themselves in China.
- [41:17]
Jasmine
You can't like, if you just use a normal VPN, like you cannot access even normalclaw.com and like, even if, like I could not access Claude when I was in China on a VPN or the Chinese Internet and it sent me a very scary email that's like, we're going to perma ban you from Claude if you like do this again. And I was like, oh my God. But the Chinese, that's okay.
- [41:35]
Nathan
You just to their office and be like, I need it back.
- [41:39]
Jasmine
That's true, that's true, that's true. But yeah, it is interesting. My even more anecdotal was that I was chatting with other friend of the pod, AJ Corabi, who now works at Anthropic, and he was recently at Land's End in SF and wearing his thinking cap, and apparently some Chinese tourist goes up to him and is like, oh, do you work at anthropic? And AJ's like, oh, yeah, I do. And then he's like, I love Claude code. I am Vibe coder. Exactly the way that you put it. And then he went on this whole thing about how he loves Claude code and he hopes to move to SF someday and all the Chinese models suck. And then he went back to China and so anyway, I thought that was a really cute story.
- [42:19]
Nathan
Yeah, I don't love just pooping on the open models because they're impressive technical achievements under resource constraint and yeah, time lag, but they're not comparable. Like the, the people love to say, like, open source is passing. And I'm like, no, what fucking universe do you fucking live in? Like, these models are just not. Like, I try each of the models every so often and it's like I could handhold this thing into something I did nine months ago that now just works and all none of the open models have given us anything like, like cloud code, GPT Pro, whatever, like, stop fucking lying. For whatever business reason that you have to fear Monger open models in China or whatever. The whole thing at Davos was really.
- [43:07]
Jasmine
Funny when I think, what was that? I mean, so much happened at Davos this year. I hate it was in a very eventful Davos that.
- [43:16]
Nathan
That caught me off guard. But I think Dario said something like, the Chinese models are relevant. Dario said a lot of things. I think he said something like, the Trident models are irrelevant.
- [43:26]
Athena
Okay.
- [43:27]
Nathan
And then the Mistral CEO was like, everybody's sleeping on the Chinese models. They're going to pass us. And it's just like, yeah, they both have their agendas there. I think obviously, like, Mistral's business model is a bit more tied to open weights because I think it's mostly on prem deployments with their own models. And Dario's known to be a bit, bit aggressive on the geopolitics.
- [43:52]
Jasmine
I mean, Dario said the H20s were nukes. Dario said we were selling nukes.
- [43:59]
Nathan
But I do think that the models are not. They're not in this conversation that we're having. Like, Xai, what's interesting was models are probably closer.
- [44:11]
Jordan
But Nathan, like, was there a moment.
- [44:13]
Kimmy
In 2025 where they were. I mean, I don't think there were times last year where this felt as inevitable as it feels right now.
- [44:22]
Nathan
DeepSeek R1 was genuinely pretty early. So like, DeepSeek R1 felt it was an excellent model and had exposed a characteristic that brought on a lot of trust, which is like seeing the reasoning change.
- [44:38]
Jasmine
It's crazy that Deep Seek was the one that like, sort of set off the see the chain of thought thing. Like, that's pretty amazing.
- [44:45]
Nathan
It really is. But I think like, the Deepseek V3 and R1 achievements were like, so high. There's some hype for the next Deepseek model. I expect it to be more in line with the other Chinese models. So if we have the likes of Deepseek's big models tend to be a lot better than QEN's big models, especially in adoption. But like GLM 4.7, the latest mini Max, like, these are very good models. I would expect the new Deepseek to beat them a little bit. But I don't think it's going to be like Deep Seek is the undisputed king of open models anymore. I think they were like, they are excellent and they are early. But it's like Kevin Ju had this post on how Deep Seek's advantage is. They don't have a business model bound to AI. I'm also like this is their disadvantage where all the other companies are existential, where it's like, we need to innovate in AI or we die. And it's like that is actually much better competitive position to be in.
- [45:48]
Kimmy
Yeah.
- [45:48]
Jordan
Nathan, as the man who himself experiences on a daily basis what not having a business model does for you, that's great.
- [45:57]
Nathan
I sell subscriptions and I work at a non profit.
- [46:01]
Jasmine
That's crazy. I can't believe you work at a non profit. Nathan. You should, you should start introducing yourself. That way when you meet people and context.
- [46:12]
Nathan
Unfortunately when I do that, they're like, what is your social cause?
- [46:19]
Jasmine
You should go to things and just be like, oh yeah, I work in nonprofits. Like I, I kind of work at an ngo. I'm really passionate about.
- [46:29]
Nathan
Accelerationism, digital literacy.
- [46:31]
Jasmine
I'm passionate about digital literacy and accessibility. Thank you very much. I want to see you do it. It'll be good. My thing is still like, are we gonna get Claud code for everything else? And I'm like, maybe, maybe not. Like, it seems much more of an open question because Claude code has all this open source code and stack overflow stuff to train on and you can verify code really fast and in automated ways and everything's written down explicitly in a code base or a lot of stuff is written down explicitly in a code base. And like, I would love Claude code for everything else. Like, I'm not precious about having the human me, like do a bunch of work. I like it when other people do my work. But like, yeah, I mean they're certainly gonna try and I hope they do. And certain types of like research things like maybe like design things like I think are gonna happen, but I don't know, I think it like. One thing that was just really notable to me is even though I've been cloud coding like crazy every day, like full on cloud code psychosis, I am absolutely not getting any more productive. And like, it's not because I'm not trying to become more productive. Like, it feels like I can only optimize around the margins, but like the core of my work is like totally untouched. And I'm.
- [47:40]
Nathan
If you wrote research papers, I think I, I think the next time I write a research paper, standard eight page CS academic paper, I'm going to fully plot a lot of it where like I can give it, I just give it raw data. It handles the plotting. I have a historical bib file with all the papers I like to cite and I'm going to have it like build out bullet points for Why I cite them and I like all the AI conferences are often boom, fuck land where they're like people can't use AI to write papers and they're all fucked. Like I don't know why they're fighting this battle but I just think like they could do an excellent job automating like helping me do the grunt work of transitioning my results into an actual paper. And I think that's kind of the boundary of that's a lot of this is because CS papers are written in something that's effectively a coding and they are self contained. Yeah, it's in a box and it's like I think that doing that well definitely at the fringe of cloud right now because it doesn't have good positional awareness and style. But I'm just going to give it a style guide. I'll be like, I always use these colors. Here are some things that I use to plot and I'm going to carry these plotting functions over with me for different things. And that's like good. It's going to be pretty good at this I think. One of my colleagues had a blog post where he's like, yeah, I need to use Claude to write help me write my meta reviews for handling conferences. It's much better at writing papers for me. And it's like there are parts of the paper that you definitely still need to write manually because that's where you crystallize the message that you're saying. But the related works is like you just have to check a bunch of boxes or battling plotting code or the formalization of the math that you already did. You need to convert it to the notation for this paper versus a different paper. It's. It's like that stuff it does really well.
- [49:28]
Jasmine
Yeah.
- [49:28]
Nathan
And I would think that eventually this can extend into some other. So it's maybe like more reporting style like clear like just like direct writing. It might help with after papers but all of these have to go through the checkboxes one after another. Fiction and like very high voice things and long term, long.
- [49:49]
Jasmine
I mean I don't even mean like fiction but even just like my like I think about my old PM job or like if I was like a product marketing person or if I was a manager or like you know, my normal like research post writing stuff like to be clear like I think AI is great and like I use it to like transcribe things and give me like when I look up words and like fact check small paragraphs and things like this and like it's awesome. I use it all the time, but it does kind of feel like I'm just optimizing around the margins all the time and like, you know, maybe I don't know when writing a CS paper, like what fraction of the work is stuff like rote work versus like thinky work. But like, I guess the thing that I feel is like, like AI is making all my rote work faster. And I'm really grateful for this and I'm really happy about it, but it is making almost no dent on the thinking, thinking work, which is how do I think that CLAUDE code will impact say the future of work or do I believe that normal people should be adopting CLAUDE code? I can obviously run a deep research report on this, but the deep research report actually doesn't help me know what I think at all.
- [50:50]
Nathan
You have a good moat. I think think. I think that is all the stuff that you guys do. That core medium I don't think you can touch. My whole thing is like I take the. Where AI is happening is a bubble. I try to skim off the front end of the bubble and then explain it. So it doesn't really matter what the AI is doing for me in terms of how I'm trying to explain it because that's just a never ending content wheel. It's like my thing is I'm trying to explain the frontier of AI and I'm trying to explain the, the where there are turns and how people train models or use models. So it's just kind of like I positioned myself there so that I will just kind of always have things to say and just where I am. But it's like I think, I guess.
- [51:34]
Jasmine
I'm just like, I think a lot of jobs are like these jobs. Like I obviously my current job is like slightly unusual but like you know, when I was a pm, like that was a very normal job. Like it would be like, oh, now make a decision of what feature we build next. And like the problem is not not knowing what potential features we could build. Like sure, maybe you know, Claude, Claude has access to my Slack and to a bunch of like SQL stuff and it can like run a bunch of, do a bunch of models and op size things and sort of like categorize people and their sentiments on the issue. But in the end like it's kind of a people persuasion like judging call game.
- [52:09]
Nathan
Like the chatbot giving you advice is probably actually more useful than cloud code. Yeah, just to be able to like talk to somebody that is smart for actually.
- [52:18]
Jasmine
But it's not agentic. Is I guess what I'm saying, like, when I think cloud code for X, I think like, is there an agentic thing? And like, like, I, I guess I'm not sure what that is for a lot of types of jobs.
- [52:27]
Nathan
Yeah, it's good for people running micro businesses as well.
- [52:31]
Jasmine
Oh yeah.
- [52:32]
Nathan
So I think that that's like, to be able to make it, it's effectively commoditizing the ability for people to build their minimal like digital home base. So there's a lot of people in the world that will benefit. It's like, it's globalizing a lot of small businesses. Sorry, did you get hooray for small businesses globalizing?
- [52:53]
Jordan
Yeah, I mean, look, we, we've talked.
- [52:55]
Kimmy
About this and it's, it's striking how like you, you think you know people, you think you live in the world and you have friends who are in different professions. But like the, the ability to actually understand like what it is that is in a job. I mean, it's, it's not surprising that all of our examples over the course of this conversation have been very personal ones because like the, if we don't have AGI that can just do everything, it is still going to be a kind of jagged thing. And I think it feels, at least for now, that we are in the universe where it will kind of fit in unique ways for different types of work. And we're not necessarily at the Dario world of anything you can do at a desktop is cooked because like AI will just be able to like run your computer and do it well.
- [53:49]
Nathan
You remember when they were talking about computer usage, like, which would like click on your screen for you.
- [53:55]
Jordan
I thought that's what the Claude cowork was supposed to be. And then they told me it wasn't.
- [53:59]
Kimmy
And then I was like, the Claude.
- [54:01]
Jasmine
Chrome extension does is actually really good at clicking stuff on your screen. I had to do like a really annoying task where I was gonna have to put in 115 separate rows through a form. And I'm sure there's a bunch of like secretaries and like white collar office workers who have to do this all day for various things. And my Claude Chrome extension browser use agent filled in from like a CSV file file, 115 lines of a form. And I was like, this is amazing. So I do think like, you know, there are a lot of people with like really boring fucking jobs who can either be thrilled that they can use the agent or who are about to be automated. I don't know which.
- [54:32]
Nathan
Some of this just takes time. To unroll like I think science broadly actually will be very accelerated by this, which is a lot of scientific institutions feel like they've kind of stalled in meaningful progress because of all the gamification of like academic institutions and stuff recently. But I do think that the drive to make scientific fields computational and those fields will kind of see massive progress just because the interface with them is often. There's often data and lots of technical literature where these types of agents can be really helpful. Then also bespoke AI in these fields I was in like, I think to be blunt, math is going to get obliterated and fully change because they're going to train models on proofs. In all like the notion of being able to prove something is going to change completely because models can prove anything. That's true by being trained on lean and stuff like this. Obviously there's things that don't work, but like math research is going to change a lot and there's like AlphaFold was a big thing. There's going to be a lot of like nuggets across science where there's these central AIs that just change fields and there's going to be a whole bunch of them over the next five to ten years time like that gives me hope. That doesn't change most people's job. I don't know what the middle ground is.
- [56:02]
Jasmine
Yeah, I do think like I see knowledge work is where it's not social. It's just like you are like go do this thing. Like that's gonna be obliterated basically.
- [56:13]
Nathan
Did you see that? OpenAI. Somebody said that OpenAI should get some sort of credit for the inventions created by OpenAI's AI models.
- [56:23]
Jasmine
Crazy.
- [56:23]
Kimmy
I read that headline like good luck with that.
- [56:26]
Jordan
The I'm sorry, you're getting $20 a.
- [56:29]
Nathan
Month and not a penny more.
- [56:30]
Jordan
The thing you mentioned earlier Nathan, about.
- [56:34]
Kimmy
The professors is like when you see the scientists and the pie charts of how they spend their time and like 60% of it is grant writing. Just like, like what you, what you talked about with those cs like paragraphs that like no one really cares about but just like have to be written. Cause it's the form factor. It will be interesting for a lot of of jobs or functions like where there is kind of dumb stuff that is just convention that it's like for whatever reason has not changed in the past 50 years. And instead if the path of least resistance is instead of reforming what a grant application looks like, you just have like the AI do it and then that.
- [57:13]
Jasmine
Well, I think the path of least resistance is that everyone uses AI to do it, and that forces the reformation. Because everyone hates reading slot. Because I know people who see grants.
- [57:21]
Nathan
And they read the. Well, the AI will read the slot.
- [57:24]
Jasmine
No, I think what happens, the good thing is that this causes them to actually reform the system is by receiving so many slop applications. Or they just choose because, like Tyler Cowen was saying for emergent ventures, like, he gets way, way, way more slop applications now, and he just, like, if he thinks it's slop, he throws it out. And so anyway, tips for anyone applying emergent ventures these days.
- [57:44]
Nathan
Put some typos in, baby.
- [57:47]
Kimmy
Oh, my God.
- [57:48]
Jasmine
Yeah, I'm actually. I think higher education is gonna, like, collapse in part because of this. And there are a bunch of these systems that, like, run on bullshit work. Like, whether that's grant applications or college applications or job applications. I think they're, like, all gonna collapse. Like, the internal contradictions are gonna come weighing down and, like, it's gonna be messy, but I think it's probably good. I think these systems suck.
- [58:08]
Nathan
Oh, yeah, I agree. I'm in the same boat. I almost feel good because I'm like a high typo person when I just write or tweet without checking it. And I'm like, like, I've got proof of human in all of my, like, cruddy tweets that are have a bad typo. And then I'm like, I gonna leave them.
- [58:28]
Kimmy
Final beat, anyone?
- [58:32]
Nathan
Another Deep Seek model will come out before we record next. I hope so. We will not be pulled into this. Yeah, I mean, like, I still. I think Deep Seek is an incredible team, but I don't like. I think the approaches other Chinese open weight labs have are gonna make them more of a. From 2027.
- [58:52]
Jasmine
The Chinese century is still ongoing. It's. It's. Yeah, it's all slowing down.
- [58:57]
Nathan
Yeah. We don't need to be sad about the state of the US on this podcast. We're happily playing our Claude video game and building things that probably don't matter in the big picture of the world.
- [59:08]
Kimmy
What?
- [59:08]
Jordan
My flower arranging app isn't going to win the century.
- [59:12]
Jasmine
I love that app. I think it's really good. I think it's such a clever idea.
- [59:15]
Nathan
You got to figure out how many people in the White House download your Flower Hour game.
- [59:24]
Jasmine
Oh, my God. Someone should teach Trump how to vibe code. I think that would be so funny. Like, I really want to know what you think of this is how.
- [59:32]
Jordan
This is how him and Dario are going to Reconcile?
- [59:34]
Nathan
Yes.
- [59:36]
Jasmine
Oh, my gosh. Trump vibe code.
- [59:38]
Kimmy
What app would Trump Vibe code?
- [59:40]
Jasmine
Well, like, I. Okay, so I'm thinking about generator. It might. No, no, I think he can do things like visualizing different buildings. Like, I could see him doing, like, Trump Sims, where it's like he take a photo of, like, a street or, like, the White House, and he's like. Tells it like, oh, make it gold. Like, put Trump's name on the top, like, da, da, da, da, da. And he could sort of like create like, a Sims utopia of, like, Trump World or something like that.
- [60:02]
Nathan
I was thinking he's gonna do a basic money compounding growth thing where he's like, how much money will I have in 48 years if I sell another crypto coin every year?
- [60:13]
Kimmy
Yeah, but him and. Him and she hanging out till 150.
- [60:19]
Jasmine
Yeah.
- [60:20]
Kimmy
All right, we're going to close this with ChatGPT and Claude's answers to what Trump would like to vibe code.
- [60:30]
Jasmine
That's a good Ethel.
- [60:31]
Kimmy
Oh, shit.
- [60:32]
Jordan
Okay, Claude. First one, Rally size AI uses satellite imagery and advanced AI to generate crowd estimates that are always three to four times what the Marshall says. Automatically generates comparison image showing his crowds versus OB almost. Oh, my God, Jordan.
- [60:51]
Jasmine
I think it's funnier than us.
- [60:53]
Jordan
A CRM that tracks everyone he's ever interacted with and assigns them a loyalty rating based on public statements.
- [61:01]
Jasmine
Claude is funnier than we are.
- [61:05]
Nathan
Chat GPTs are going to be bad. Did you ask ChatGPT?
- [61:07]
Jordan
ChatGPT? Still thinking. Let's. Let's give Kimmy a shot. Kimmy always. Kimmy always gives me something kind of clever.
- [61:14]
Nathan
You can ask Rock while you're at it.
- [61:16]
Jordan
Okay, you do that.
- [61:17]
Nathan
My Twitter's blocked for the day.
- [61:19]
Jasmine
My Twitter is also blocked right now.
- [61:21]
Kimmy
Oh, my God.
- [61:21]
Jordan
Tariff Master 3000. A calculator where you input a country and it outputs a tariff percentage. The Math is just math. Dot random times 100 +50. It has a reciprocal button that doesn't actually do math, but just adds 25%.
- [61:39]
Kimmy
No.
- [61:40]
Nathan
Frog is currently searching the web. It read 40x post. No, it would be Trump Script. The programming language itself. Syntax, tremendous error messages. Pure Trump quotes. Sad. You forgot a semicolon. Nobody forgets semicolons like Sleepy Joe. Every program must end with Make America Great Again or it won't compile. Features include Build the Wall and Rigged election Check.
- [62:07]
Jasmine
Okay. I think there's, like, a great side project bit to be done that's like making Trump's GitHub like, you know, it's like the JMail project. But it's like we make Trump gith and it has all of these in it. This is our new vibe coding project.
- [62:19]
Nathan
Guys, I'm out.
- [62:22]
Jordan
Okay, chatgpt is not funny. Kimmy's also not funny at all. Okay, Claude, you have the mandate. Congratulations, Dario. You did it. Hats off to you. Excited for your super bowl appearance?
- [62:40]
Kimmy
You got my $200 a month. Everything's turning up orange. I don't know what to tell you. It's a good model.
- [62:49]
Jasmine
It's a good model. Okay.
- [62:52]
Nathan
Happy New Year.
- [62:53]
Jasmine
Happy New Year. It's too late to say that. Happy cloud coding.
- [62:56]
Nathan
Happy Feb. Happy Valentine's Day.
- [62:59]
Jordan
Happy Valentine's Day.
- [63:00]
Jasmine
Oh, that's sweet. Thanks, Nathan.
- [63:02]
Jordan
Please, everyone, everyone, use. Use Jordan's flower aging app. Get your loved one a digital Monet inspired.
- [63:10]
Jasmine
Wait. There's gonna be so many vibe coded Valentine's Day software programs. I think that's gonna be cute. I think. I mean, for the right person. It's gonna be adorable. We gotta do. We gotta ask Claude what Xi Jinping would vibe code.
- [63:23]
Kimmy
Okay. Okay.
- [63:26]
Nathan
Okay.
- [63:26]
Kimmy
That was perfect. What would Xi Jinping and Putin vibe.
- [63:36]
Nathan
Together?
- [63:37]
Kimmy
Separate and together or.
- [63:39]
Nathan
Okay.
- [63:41]
Kimmy
I am so impressed with those Claude ones.
- [63:43]
Jasmine
That was really good. That was amazing.
- [63:45]
Nathan
I think Claude's quietly become like. It kind of has the spunk.
- [63:50]
Jasmine
Yeah.
- [63:51]
Jordan
Okay. Xi Jinping Common prosperity calculator. You input a tech billionaire's net worth and public profile and it outputs exactly how much they should voluntarily donate and whether they need to disappear for a few minutes. Months of personal reflection. Century of humiliation. Reminder. A daily notification app for Chinese diplomats that automatically generates historically aggrieved talking points. Today in 1842, for instance. And now for Putin. We've got to be rough. Long table augmented reality that automatically previews how small visiting leaders will look from him 40ft away. Historical Russia.
- [64:35]
Kimmy
I.
- [64:36]
Jordan
You draw any border on a map that generates a 3,000 word justification for why that territory is historically function.
- [64:44]
Nathan
Oh, my God.
- [64:45]
Jordan
Oligarch manager or portfolio tracker. But for oligarchs, track their yacht via satellite monitors for disloyalty signals. Has a one click widow alert feature.
- [64:55]
Jasmine
Jesus Christ.
- [64:57]
Jordan
Okay. Tucker Booker. An app for scheduling friendly interviews with contrary Western media figures.
- [65:06]
Nathan
No. What if I put China talking.
- [65:10]
Jasmine
I want to put Nyahu next, but I think we gotta wrap up.
- [65:14]
Nathan
We could do it. We could do it. We could do it.
- [65:16]
Kimmy
Okay.
- [65:17]
Nathan
No one's listening at this point. It's okay.
- [65:20]
Jasmine
No, I think we gotta clip this and put it at the front. Jordan. I think we have a mom.
- [65:23]
Jordan
Donnie.
- [65:24]
Jasmine
Oh, My gosh. Yeah. Okay, while you figure this out, I'm gonna open the door for the person who has arrived 15 minutes ago. Okay.
- [65:31]
Nathan
Let him in. She doesn't have Bluetooth headphones. She's lost. Okay.
- [65:39]
Kimmy
These are not that funny.
- [65:41]
Jordan
The best one is. Yeah, your post review. Intercepts his son's tweets before they go live, flags anything that will cause an international incident. Sarah Scheduler, which is his wife, who's.
- [65:53]
Kimmy
Like, corrupt and terrible, manages household staff.
- [65:55]
Jordan
Rotations, tracks which ones might testify, and reminds him to order catering from approved vendors only.
- [66:01]
Kimmy
It's a deep. Guys, I did the Netanyahu jokes, which they're.
- [66:07]
Nathan
Man, they're not.
- [66:08]
Kimmy
People wouldn't find that funny.
- [66:11]
Nathan
I'm gonna cut it.
- [66:12]
Jordan
Okay. Mom Donnie.
- [66:14]
Nathan
Wait, this is.
- [66:14]
Jordan
This is Mahmoud.
- [66:15]
Kimmy
Mom Donnie. No. Like, NYC Mom Donnie.
- [66:19]
Nathan
Claude's gonna be like, I don't know who that is.
- [66:22]
Jasmine
Yeah, I wonder if Chachi PT or is any less funny because it has, like, you know, training about political bias.
- [66:28]
Athena
Or something like that.
- [66:29]
Nathan
That maybe.
- [66:31]
Kimmy
Yeah, all the. All the chat GBT ones are very, like, very prosaic.
- [66:35]
Nathan
Like.
- [66:35]
Kimmy
Like one tap clip Factory rally ops, merch drop machine, deal tracker, CRM. I mean, this is. This is embarrassing.
- [66:43]
Nathan
Thanks for listening to Overwatch over.
- [66:45]
Jasmine
Not Overwatch.
- [66:46]
Kimmy
Who has. I.
- [66:47]
Nathan
You've been gaming.
- [66:48]
Kimmy
Hey, man, I'm.
- [66:50]
Jordan
I'm free tonight.
- [66:51]
Kimmy
I got. I got babysitting duty.
- [66:53]
Jasmine
Oh, my gosh.
- [66:54]
Athena
Okay.
- [66:54]
Jordan
I want to play some Overwatch nature.
- [66:56]
Nathan
I've never played.
- [66:58]
Jasmine
That was fun.
- [66:59]
Kimmy
We'll call it here.
- [67:01]
Jasmine
Bye, guys.
- [67:01]
Dario
Here's what Dario thinks.
- [67:03]
Nathan
Trump will vibe Code Loyalty Score.
- [67:07]
Dario
A CRM that tracks everyone he's ever interacted with and assigns them a loyalty rating based on their public statements. Sends him push notification whenever someone on cable news says something nice or critical. Tariff Master 3000. A calculator where you input a country and it outputs a tariff percentage. The algorithm is just mad random. 100/50 has a reciprocal button that doesn't actually do math, just adds 25%. Now for C Boxing. Automatically checks any documents, speech, or WeChat message against each and ping thought and highlights ideological deviations. Has a study score leaderboard for politburo members. Anyone below 95% gets a calendar and invite title T Common Prosperity calculator. You input a tech billionaires net worth and public profile file, and it outputs exactly how much they should voluntarily donate and whether they need to disappear for a few months of personal reflection. Pooh Shield. An image recognition app that scans the entire Chinese Internet in real time and scrubs any Winnie the Pooh Comparisons has deleted 4.7 billion images. Century of Humiliation Remember a daily push notification app for Chinese diplomats that generate historically aggrieved talking points today in 1842 with suggested wolf warrior responses. Now for Putin long to learn augmented reality app that lets him preview exactly how small visiting leaders will look from 40ft away. Has presents for Macron and Tucker is or Russia Eye. You draw any border on a map and it generates a 3,000 word historical justification for why that territory is historically Russian. Cites obscure 12th century treaties. Oligarch Manager a portfolio tracker but for oligarchs, tracks their yachts via satellite, monitors for disloyalty signals and has a one one click window alert feature. Now for Zorin free the bus account down to fare three implementation. That poke he'll keeps resetting has a Albany said no again notification sound that's just a sad trombone socialist but make it work. An app that translates DSA platform planks into things the city can actually do without state approval. Output is always very short.